AutoFigure

AutoFigure is an intelligent system that leverages large language models (LLMs) combined with iterative refinement to generate high-quality scientific illustrations from text descriptions or research papers. The project has been accepted by ICLR 2026.

1. Purpose Automatically transform text descriptions or research papers into scientific figures ready for publication. It ensures high accuracy and aesthetic quality through a Review-Refine loop, whose process includes: generation (the agent creates an initial SVG/XML based on the description and reference materials), evaluation (the critic gives a quality score from 0-10 and provides specific feedback), and refinement (the loop continues until the figure meets publication standards).

2. Main Capabilities

3. Extract methodology from papers and automatically generate figures.

4. Generate multiple figure variants with enhanced aesthetics.

5. Provide FigureBench: the first large-scale benchmark for generating scientific illustrations from long text.

6. Related sister projects and version updates:

• [2026.03.24] DeepScientist v1.5 is released, a local-first open-source autonomous research system for end-to-end scientific discovery; explore it on GitHub or read the ICLR 2026 paper.
• [2026.03.11] The AutoFigure-Edit paper has been published on arXiv and featured on Hugging Face Daily Papers.
• [2026.02.17] The AutoFigure-Edit online platform is launched, free for all scholars to use. Try it at deepscientist.cc, or check out the open-source code on GitHub; this Edit version performs better.
• [2026.01.26] AutoFigure is accepted by ICLR 2026; read the paper on arXiv.

1. How to Use Two ways to use it are provided:
2. Python SDK (recommended): install by cloning the repository.
3. Web Interface: suitable for visual interaction and editing.

In terms of configuration, AutoFigure is highly configurable and can be set in Config() or through environment variables. Among the supported LLM providers, the Bianxie AI aggregated API provides users in mainland China with friendly access to AutoFigure's recommended Gemini models without the need for an overseas credit card, and you can register at bianxieai.

API Reference: generate() parameters; generate_from_paper() parameters accept all parameters of generate(). The result object is GenerationResult; Enhancement Modes are also provided.

Project structure, community, and support: A WeChat discussion group is provided. If the QR code expires, you can add WeChat ID nauhcutnil or contact tuchuan@mail.hfut.edu.cn; the friend request note format is "University/Company - Degree/Position - Name".

IV. Applicable scenarios Suitable for researchers and scholars to quickly generate publication-grade scientific illustrations when writing papers and reports; also suitable for scenarios that require automatically extracting methodology from papers and visualizing it; suitable for users who want to visualize, interact with, and edit charts on the web.

V. Limitations and notes

1. It is recommended to use Google AI Studio's gemini-3.1-flash-image-preview as the image generation model and gemini-3.1-pro-preview as the text model; each run costs about $0.50, consumes about 30,000 tokens, and takes about 20 minutes.
2. Gemini's terms of service do not allow users in mainland China to access or use it. If OpenRouter reports an error, it is often because accounts registered in mainland China lack the permissions required to use Gemini; it is recommended to use an OpenRouter account registered in the United States or Europe and ensure compliant use.
3. The project is open source under the MIT License, see LICENSE for details; the use of the name and logo is additionally explained in TRADEMARK.md.